reacharavindh

One more item on my todo list likely got a booster.. instead of building it from scratch I might build mine on top of this.

Seriously, the subscription frontier models scratch my control freak nerves and lets me build my computing set up just the way I want and not a single bit more. I’m currently making a minimal Linux distribution for myself based on Gentoo with the idea to develop any and all graphical applications except Firefox _for_ myself using LLMs. Nothing on my system I don’t need.. It sounds absurd but is still possible in 2026 with these subsidised LLMs. They are not my private data, so I don’t mind having the agents build it, test it, and automate it. If they want to train on my calculator app and CAD built upon other code from the Internet they are already trained on, they are welcome to :-)

This right here is a new age of personal computing.

My guess would have been that every tenant gets their own “handler/driver” model that updates its weights based on the user intent, problem etc.. but it has “experts” to call on demand that are pre-trained set weights. I thought this is how MOE models worked. Then a user gets access to large weights where it matters..

It isn’t just winner takes all anymore(hasn’t been for years?)… I don’t think any reasonable consumer/company would hitch their wagon to one AI provider unless they came with an order of magnitude better quality at competitive costs.. For example, we use Claude Opus 5.5 for particular tasks that justifies its cost, but use self-hosted(not really… hosted on AWS Bedrock, but the point applies in spirit) cheaper models from Anthropic’s competitors and even open weight models for specialized tasks that don’t need the frontier model. Even if Anthropic pulls a huge lead against their competitors, there is no reality where the whole world sticks to Anthropic and shuns cheaper models. This is why their absurd valuation sucks. At reasonable valuation, they are quite a bet because Opus 5.5 is still a great workhorse as of today.

I think Deepseek did this in the most focussed way.. their own optimization (key value caches, downscaled images for vision support etc) in their inference stack to push the boundaries in favor at a time of hardware scarcity.

It feels like OpenAI and Anthropic have a ton of room to optimizing that regard, but they choose to rather focus on moving the frontier forward than doing things that are already possible in a more efficient way.

What is this story? Do you have any references on what happened? What did Claude code start to use gzip for?

Indeed they're non-technical problems, but really blocking ones. As a user, I dont want to have my filesystem and Kernel have a possibility of going out of sync leading to an unusable system during a regular update.

The biggest attractor towards BcacheFS for me was the possibility of that cohesion. All the technical merits were a bonus as ZFS does already what I wish for. That factor went away with BcacheFS choosing to not get mainlined into Linux Kernel. Now we have two good filesystems outside the kernel :-(

My layman's understanding is that these models are very capable pattern matchers. When they need to do X, they go on a hunt for every pattern they could find and try to satisfy solving X. If that is not deemed good enough? make them dig harder and they take a different route to find patterns to solve X. Now, their arsenal is everything that they have been trained on(which is a lot!). This ability comes in extremely handy to apply existing techniques and patterns in ways we humans didn't quite think of.. because we are domain driven. A mathematical model about finding a good enough approximation of something at much lower cost is perhaps well understood in its applications in mathematics or whoever got to see those results. But, these LLMs, when pushed harder and harder might bring that to apply and solve an entirely different problem like for example, classifying whether some event is fraudulent. When it does such things, it _feels_ novel because it ended up using something we haven't thought of yet. But, it got there by brute force, stroke of luck, or some random path. I guess this is the reason why the frontier labs run those crazy swarms of AI agents to uncover all such paths turn them into chartered thinking vectors instead of random attempts.

I like to believe we humans still retain the ability to think completely outside the box and discover new phenomena and understanding aided by these pattern machines. In other words, these models are great at running combinations through patterns and we humans go one level higher and add those new patterns for them to accelerate us even more. We determine what to solve and the LLMs help us solve it faster.

Most important detail…. It was prior literature that documented the possibility of doing this and the systems fixed those vulnerabilities as well. The AI model probed their current infrastructure and found an unplugged but pre-documented(therefore on their training) hole.

This detail matters because the idea itself is not novel. LLMs did cleverly discover, filter and use the patterns.

Watch Greg Koah’s recent talk on AI discovered vulnerabilities in Linux kernel to understand the limitations of LLMs aided offensive security acts. It was a rational light for those who fear monger that AI will _just_ take over humanity and be intelligent in novel ways.

The larger threat still remains though - if any literature anywhere threw the scattered dots about techniques, issues, rough ideas etc that were not acted on yet.. these pattern machines are going to find them and use them to achieve their goals(if that were to break through). Just not in the sentient way folks assume based on news.

I understand why Google would run Gmail for free.. it’s signals for their ad machine even if they don’t literally push the contents of the emails themselves directly. I mean the metadata, connections etc is valuable for ads intelligence.

But I toss my otherwise public git repos on GitHub.. if they weren’t handing out for free, I’d have hosted them anywhere else publicly as well. What do they get by hosting my _private_ repos for free? Did they explicitly say anywhere that they are training models on code in private repos? That’ll be a dark pattern hard to miss out here on HN!

GitHub lets me create public and private repos.. with a free account.. how are they making money?! Why are they throwing hardware at this explosive growth of repos with LLMs?

I was thinking the economics of it the same way.. if the AI companies are burning debt and VC money to give me cheaper tokens why not take them for it? I can always rent a GPU server for those personal, private AI needs.. heck even run a decision model to route them to the rented GPU servers either on demand or scheduled for few hours in a day…

It doesn’t have to be black or white…

For my toy projects, adventures, learning exercises, and throwaway MVPs I will prefer the subscription plans… they are cheap for what they offer today.

N you please answer one thing for my peace of mind?

If I use a TV like the ones you’ve worked on, and never ever connect it to any WiFi network since the day I bought it, is there any way they still call home and pass their payload? Like searching for and connecting to an open WiFi network, some kind of relay through another device over Bluetooth or any other way I haven’t thought of yet?

It’s a pity we gave in to the “ecosystem” play by these huge companies.. even if I jailbreak my iOS and never let it update.. Apple can easily detect the jailbreak and lock me out of anything iCloud, and iMessage and such. iPhone would be one glorious iPod touch after that.. :-(

Seeing all these gives me the same energy of early days of jailbroken Android phones where one could become root and do whatever with the hardware… I stopped considering such tinkering because I got tired of finding those little tweaks and fixes on forums like xda every time something was updated. With the LLMs capable of automating that tool, I might find those old school way of mobile phones again…

The line has always been blurry..

Why use power tools(planer vs hand planes, power saw vs hand saws, saws instead of chisels etc) when you could do things by hand?

It’s all about which parts one enjoys doing by their own and which ones they’re willing to abstract away with a tool.

Some folks abstract away the whole process and just visit IKEA :-)

> a personal tendency to try and make it pixel perfect and because of that I sink way more time into it than I should.

Now that resonates too close to home...

May be I will come to the same conclusion as you after spending hours/days/weeks into trying out a bunch of things with CAD. Lets see... this rabbit hole feels worthy of jumping into.... (it will be different this time!)

I understand the sentiment... As a computer engineer who enjoys woodworking because it is tangible, abstraction-free, and in full control right in front of your eyes, the line I draw is not "I won't allow computers into this Zen wood shop". Instead, I use computers to accelerate the error prone parts of human planning, and also expand my cognition by making things visual beyond simple imagination. If I see that Coffee table proportions I had in mind as a CAD model, I may spot any balance issues, and even validate if that is what I was going for aesthetically. Heck, it might even give me a chance to improvise on the design because I have now visualised it outside my mind.

Once the CAD model validates my idea, and gives me a rough plan, the zen part of working with the wood, adjusting for grain, knots etc can begin. Even while making changes on the fly, a visual model of the original plan comes in handy.

I tried this for my workbench a few years ago.

https://aravindh.net/wood.html#gallery-7

https://aravindh.net/wood.html#gallery-8

and it significantly helped without removing the joy of making mistakes..

OP here. writing for my own reference. Found great pointers so far.

I'm going to try https://sawdust.diy and https://fluidcad.io/ for the short term.

For the long term, I have noted FreeCAD + Blender + $ai_model to be a good general purpose workflow that yields more control, and will probably get me lost into the details.

https://janne-alatalo.fi/en/articles/silhouette-modeling/ and its accompanying SKILL - https://github.com/janne-alatalo/image-to-cad-skill

FreeCAD MCP: https://github.com/neka-nat/freecad-mcp

Woodworking specific FreeCAD resources : https://github.com/dprojects/Woodworking

Two YouTube Channels related to FreeCAD + woodworking :

1. https://www.youtube.com/@dprojects.woodworking

2. https://www.youtube.com/@ga3d_._tech528

There goes the weekend! Cheers HN!

Hey. This is very cool. I will definitely try this out this weekend for the projects I have in progress.

My personal experience is that woodworking at hobby grade is a art+science+engineering in varying proportions depending on the person's interests. I can relate to your friend in that almost always, as I work through a project, it evolves in its details, joinery, design (even size as I make my mistakes) etc. However, it is important to visualise _what_ one wants in mind before beginning to mill wood. I sketch so much in various perspectives to narrow that down and try to visualise how my idea would materialise.

If AI could catalyse that first part and help visualise the idea in many perspectives that are easy to tweak and improve, I think it will be a solid base to work upon and change organically as we build it in the shop.

This looks very nice. I can see how opinionated it is in a way that it suits your needs well. Cheers!

As an engineer obsessed with the finer details and getting his fix from weekend woodworking by making things just the way I want for fun, and tinkering with newer ways of doing things, some finer details that probably won't make a lot of rational sense, but I really badly want them because it is cool.. I want a more general purpose workflow where I can make anything I feel like... {}toys for kids, bike stand for the shed, A tool wall, Bed for the kids..}. Your approach to doing what you happened to need - bookshelves, workbench etc in a customisable way is very cool. But, I've already made those over the years and need something else.. :-)

Your work gave me a lot of ideas and levelled up what output I can make once I figure out a more general purpose workflow.

I would like one like this.. (Slack/Discord alternative) with a native iOS app, realtime push notifications, and self-hosted backend such that my AI assistant can be a “user” on it and participate in conversations like a team member(who has a lot of knowledge at their finger tips).

I can’t find any such free thing that has a native iOS app.

Buzz was one I looked at but it was way too complicated with Nostr and a bunch of things that are way too much for what I intend to use it for… :-(

This - Strafe, Chatto, Stoat all fall apart in the native iOS app requirement..

List is the same as the incumbent model :-( I was hoping for a lossless Atmos TrueHD pass through support.

This is bait and switch..

You can’t just use your Claude sub on a third party harness without getting blocked even if you use the Agent SDK. Apparently, Anthropic gives “credits” depending on your sub and you can use those _credits_ when you use third party harnesses or Claude -p mode. That usage is not the same Claude sub one would assume it is… It is all quite deceiving. Don’t get a Claude sub unless you’re okay with being locked into Claude Code…

Is BCHS still maintained?

The last update seems > 5 years ago.

https://github.com/kristapsdz/bchs

I may need to take a second look at Pi again if it has any niceties like that notable boost in round trip performance compared to hax.

I mostly stuck around with Hax because of a few design choices. The sessions are merely plain text files indexed by a dirpath slug. This means I can just start a hax session in that directory and expect sessions history to be available in any model.. (I occasionally dabble in GLM for security checks, Gemini for writing stuff, Deepseek for tightly scoped tests etc).

I’ve been using GPT models via hax[1], a bare bones harness with minimal tools on my system and have the models rely on bash for the most part the last few months. It’s been a breath of fresh air amidst all these overengineered harnesses.

If I need special actions, I just make them CLI tools on my Mac and tell the AI model that such a tool exists(via a single line in its system prompt) and that’s it. It’s been very nice.

[1] - https://usehax.dev/

No affiliation whatsoever. Just a happy user.

I’d usually follow that logic that using AI models(that basically trained on everything it could get hands on without regard for people’s labor and effort) to steal someone’s right to get paid for their effort. But, this is Adobe we are talking about…

1. Nobody asked them to bundle PhotoShop with all the cloud features and force it down their user’s throat.. Adobe Photoshop as a tool could easily be a one-time paid per version software. Want assets and things only possible in the cloud? Here is an optional cloud feature that costs $x should be the logical default.

Force your users to pay for stuff they didn’t ask for and lock them them in for years? It feels morally justified that they get wrecked by AI models replicating their tool. If they are legitimately that good that one needs their Cloud justified content, perhaps they will still pay for Adobe. But, that’s all they are entitled to at this point.

This is the first time I am hearing of yggdrasil. I currently have a similar setup using Tailscale. I always thought Wireguard was teh primitive here and Tailscale just made it more convenient in the pointy clicky way. Why Yggdrasil instead of Wireguard which at the outset sounds more standard solution for this need?

The developer may be making extreme claims. But, there is a sliver of cold truth hiding behind it.

The LLMs are so adept at pattern matching and combining so many bits and pieces of engineering knowledge that they can now easily take over this aspect of software engineering- “using existing building blocks to build out a tool or service”.

A liberation of software tools and services. See something you like? Build it for yourself just the way you like it. There is no moat in tools and software anymore. The new reality whether one likes it or not is that anybody can replicate your software and they most likely will. Not even the “assets” you guard and let only _your_ services access because LLMs can scrape relentlessly, fill-in the gaps and get good enough to be a viable alternative.

The only moat left are in hardware, hardware adjacent software, and services that wield genuine network effects to be useful.

For other HN folks(or bots) passing by... KMP = Kotlin Multiplatform Apparently, unlike Electron, it lets devs write shared Kotlin code once and compile/use it on multiple platforms. KMP can compile Kotlin to native binaries for platforms such as iOS..

Interesting. TIL about this. I added myself to the waitlist.